Classifications

• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
  • H04N23/10—Cameras or camera modules comprising electronic image sensors; Control thereof for generating image signals from different wavelengths
  • H04N23/12—Cameras or camera modules comprising electronic image sensors; Control thereof for generating image signals from different wavelengths with one sensor only
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N25/00—Circuitry of solid-state image sensors [SSIS]; Control thereof

Definitions

• Embodiments of the present invention relate to systems and methods for compressive sampling in imaging.
• Images herein can be considered analog signals whose amplitude may represent some optical property such as intensity, color and polarization which may vary spatially but not significantly temporally during the relevant measurement period.
• Light intensity typically is detected by photosensitive sensor elements.
• Conventional image sensors are typically composed of a two dimensional regular tiling of these individual sensor elements.
• Color imaging systems need to sample the image in at least three basic colors to synthesize a color image. We use the term "basic colors" to refer to primary colors, secondary colors or any suitably selected set of colors that form the color space in which the imaging system represents the image.
• Color sensing may be achieved by a variety of means such as, for example, (a) splitting the image into three identical copies, separately filtering each into the basic colors, and sensing each of them using separate image sensors, or (b) using a rotating filter disk to transmit images filtered in each of the basic colors in turn onto the same image sensor.
• CFA color filter array
• demosaicing aka demosaicking
• Conventional demosaicing algorithms typically involve the use of, for example, interpolation techniques such as bilinear, demodulation and filtering and edge adaptive algorithms.
• Conventional demosaicing algorithms work well only if the high frequencies, corresponding to the fine detail, of images in the basic colors are correlated or have low high frequency energy content in at least one direction.
• Image compression is typically applied after digital image acquisition to enable reduction of the system data load during transmission and storage.
• Image compression is based on the observation that natural images and many synthetic ones are approximately sparse in some basis. This includes the Fourier related bases, for example, the discrete cosine transform ("DCT"), employed by JPEG and wavelets which rely on empirically observed hierarchical self similarity of natural images and underlies the JPEG2000 compression method.
• DCT discrete cosine transform
• compressive sensing A new reduced rate sampling scheme called “compressive sensing” has been developed recently.
• the goal of compressive sensing reconstruction techniques is the solution of ill-posed inverse problems through the regularization scheme known as "sparsity promotion.”
• Ill-posed inverse problems here are concerned with reconstructing an original signal from a sampled data set of a transform of that signal, where the transform is non-invertible.
• Sparsity promotion uses prior statistical knowledge of the original signal's sparsity in some basis to search preferentially for solutions of ill-posed inverse problems that are also approximately sparse in that basis. See Candes et al., "Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information" (IEEE Trans, on Information Theory, 52(2) pp.
• Incoherence is a measure of dissimilarity between the two bases; more precisely, it is the largest magnitude of the inner product between any pair of basis vectors from the two respective bases. See Candes and Romberg, Sparsity and incoherence in compressive sampling. (Inverse Problems, 23(3) pp. 969- 985, 2007). They derived an inverse relationship between the incoherence between the bases and the number of samples required to accurately reconstruct the original signal with high probability. Compressive sensing techniques thus reconstruct the original signal from an under- determined system of equations through a joint maximization of logical tenability and physical probability.
• x 112 Equation (1) where x is the solution in the sparse basis, rasterized into a single dimensioned vector, y is the observed image, also rasterized and A is the transformation matrix representing the change in basis from the sparse to the spatial domain, x, being the sparse representation, has fewer elements than y.
• the first term penalizes deviation from the observed data whereas the second term is an Li norm that has been shown to penalize less sparse solutions, ⁇ controls the relative weights of the two penalty terms.
• Constrained convex optimization problem formulations also exist which minimize just the first or second term while constraining the other below a threshold.
• Orthogonal matching pursuit and its many variations such as Simultaneous Orthogonal Matching Pursuit, Staged Matching Pursuit, Expander Matching Pursuit, Sparse Matching Pursuit and Sequential Sparse Matching Pursuit form a popular class of algorithms that obtain an approximate solution quickly.
• Total variation TV
• TV Total variation
• Yet another class of reconstruction algorithm is motivated by de- noising methods and includes iterated thresholding in a transform domain. Subsequent developments continue to further improve reconstruction quality and reduce the computational burden.
• Compressive sensing is information scalable, i.e., even if too few samples exist to do an exact reconstruction, various levels of information can be extracted depending on the number of measurements.
• compression sensing means reconstructing a signal using prior statistical knowledge of the original signal's approximate sparsity in some basis to search preferentially for solutions of an ill-posed inverse problem, based on samples of a transform of the original signal, that are also approximately sparse in that basis.
• GPSR This solves a bound-constrained quadratic programming formulation using gradient projection algorithms. It is described in Figueiredo, Nowak, Wright, “Gradient projection for sparse reconstruction: Application to compressed sensing and other inverse problems" (IEEE Journal on selected topics in Signal Processing, 2007), hereby incorporated by reference in its entirety, and is currently available at http://www.lx.it.pt/ ⁇ mtf/GPSPv/.
• 11-Magic This solves a "linear programming" or second-order "cone programming" formulation. It is currently available at http : //www . acm . caltech . edu/'l 1 ma gic/ or may be obtained by request from the author Emmanuel Candes, Stanford University Department of Mathematics, 450 Serra Mall, Bldg. 380.
• ll_ls This solves the convex quadratic programming formulation of equation (1) using interior-point methods. It is currently available at http://www.stanford.edu/ ⁇ boyd/ll_ls/. Compressive Imaging
• DMD digital mirror device
• the present invention provides a method for processing an image.
• a sample set of data generated by transforming and sampling an optical property of an original image in a spatial basis, wherein the transformation effected is substantially diagonal in the spatial basis is received.
• a compressive sensing reconstruction technique is applied to the sample set of data to produce a set of inferred original image data.
• an image sensor having a plurality of photosensitive sensor elements may be provided, along with an optical transform having a plurality of transformation elements responsive to the optical property, wherein the transformation effected by the optical transform is substantially diagonal in the spatial basis.
• the image is projected onto the optical transform, and the optical responses of the transformation elements are sensed at the image sensor to the original image to produce the sample set of data.
• the present invention also provides a system for processing an image.
• the system comprises a data processing apparatus adapted to receive a sample set of data generated by transforming and sampling an optical property of an original image in a spatial basis, wherein the transformation effected is substantially diagonal in the spatial basis. It then applies a compressive sensing reconstruction technique to the sample set of data to produce a set of inferred original image data.
• the system may further comprise an image sensor having a plurality of photosensitive sensor elements; an optical transformation device having a plurality of transformation elements responsive to the optical property wherein the transformation effected by the optical transform is substantially diagonal in the spatial basis; and an optical imaging device adapted to project the image onto the optical transformation device, the image sensor being disposed with respect to the optical transform so as to receive the optical responses of the transformation elements to the image at the photosensitive elements corresponding thereto to produce the sample set of data.
• One preferred embodiment of the present invention combines a color filter array with colors arranged randomly or pseudo-randomly with an image reconstruction technique based on compressive sensing that achieves the same limiting image resolution in each color as a monochrome image sensor using the same number of sensor elements.
• FIG. 1 is a flowchart showing a method for sampling and reconstructing a color image using compressive sampling in accordance with a preferred embodiment of the present invention.
• FIG. 2 is a schematic diagram of a color imaging system in accordance with a preferred embodiment of the present invention.
• FIG. 3 is a diagram of a color filter array in accordance with the preferred embodiment of the present invention.
• FIG. 4 shows the results of a simulation of color image sampling and recontruction using compressive sampling in accordance with the preferred embodiment of the present invention as carried out in Matlab.
• FIG. 5 lists a portion of Matlab code for producing a random color filter array suitable for use in a preferred embodiment of the present invention.
• the present invention uses the spatial basis directly as the sampling basis.
• This direct spatial sampling is distinct from the previously mentioned existing techniques of image sampling because, among other things, it samples the image after applying a transformation, such as color filtering, that is substantially diagonal in the spatial basis, i.e., the transformed value at each location depends on the original image at that location only, not at locations substantially removed from it to achieve incoherence between the measurement and sparse bases.
• a transformation such as color filtering
• optical elements that affect transformations that are non- diagonal in the spatial basis such as optical low pass filters are allowed in the pipeline, but the original image- the image we aim to reconstruct using compressed sensing techniques - is defined herein as the image produced after such filters have been applied. Moreover such elements are not used for the purpose of incoherence enhancement.
• the naive design of choosing a sensor array with fewer sensor elements arranged in a regular pattern does not offer the advantages of compressive sensing: the maximum spatial frequencies of such image sensors is limited by their reduced Nyquist frequencies. Frequencies beyond the Nyquist limit cannot be inferred since the random projections of these frequencies in the spatial measurement basis cannot be distinguished from those of their aliases.
• a regular monochrome image sensor with a few randomly or pseudo-randomly chosen sensor elements turned off can still provide enough information to reconstruct the image at full resolution with little or no loss of quality. While this does not lead to a reduction in sensor area, it does reduce the read-out rate. This forms an embodiment of the present invention.
• the preferred embodiment of the present invention uses the sparse sensor elements in the monochrome image sensor described above to pack three basic colors into an image sensor with the same number of sensor elements as a monochrome image sensor and still obtain the same limiting resolution in each color as the monochrome sensor. Approximately equal numbers of filter elements in each of the basic colors, arranged randomly are used to multiplex the basic colors.
• Another embodiment uses panchromatic filter elements wherein each element is transmissive to varying degrees in all basic colors.
• the color compositions, not just their distribution are randomly chosen.
• CFAs are harder to manufacture than CFAs in basic colors.
• Another embodiment uses filters of basic colors but with different transmittivity, or augments the CFA with a few non-basic colors or clear filters. This is done to improve sensitivity or the dynamic range of the resulting sensor.
• FIG. 1 A flowchart showing a preferred method of compressive imaging is shown in FIG. 1.
• a CFA is provided with colors arranged randomly or pseudo-randomly.
• the term "randomly” is to be construed as including psuedo randomly as well as raondomly.
• the incident image is filtered through this CFA.
• the filtered image is detected by an image sensor in a single measurement period.
• the full resolution image is reconstructed from the image sensor output and the CFA pattern using compressive sensing sparse signal reconstruction techniques.
• FIG. 2 A schematic diagram of an imaging system in accordance with the preferred embodiment of the present invention is shown in FIG. 2.
• Image 210 is focused by lens 220 onto an optical transformation device, specifically CFA 230 (As used herein, the term "optical transformation device” means a device, such as a fixed color filter or a spatial light modulator that transforms an optical image from one state to another based on one or more optical properties of the image.).
• the filtered image is detected by image sensor 240.
• the resulting plurality of sensed filtered image intensity values is sent to processor 250, which implements the algorithm of FIG. 1, where full resolution image reconstruction is performed.
• FIG. 3 An exemplary CFA 310 in accordance with a preferred embodiment of the present invention is shown in FIG. 3.
• red, green and blue (collectively "RGB”) filters in approximately equal numbers are distributed in a randomized pattern.
• a Matlab simulation was performed wherein a CFA with approximately equal numbers of red, green and blue filters arranged in a randomized pattern was generated. Matlab Code for generation of such a CFA is shown in FIG 5. The original color image was filtered through this CFA, a certain amount of white noise was added to the filtered image and reconstruction was performed on the resultant image sensor output. The 2D discrete cosine transform ("DCT") basis was used as the sparse basis.
• the GPSR solver was used to perform signal reconstruction. The GPSR solver solves the formulation of Equation (1), shown above, by converting it into an equivalent bound-constrained quadratic programming formulation and then using gradient projection algorithms to solve it, as will be understood by a person having ordinary skill in the art .
• FIG.4 shows the results of this simulation.
• Image 410 is the original color image.
• Image 420 is the original image after being filtered by an exemplary randomized RGB filter.
• Image 430 is the reconstructed full resolution color image. The color images are provided here in grey scale.
• FIG.5 lists Matlab code used in the simulation of a simple embodiment of the present invention.
• the Matlab code for generating the CFA should be sufficient for anyone of ordinary skill in the art to reproduce these results using the GPSR solver.
• Matlab is a product of The Math Works, Inc., Natick, Mas sachus setts, U.S.A.
• the present compressive sensing imaging design enables integration of compression in the image acquisition layer. This reduces the data transfer requirements off of the focal plane - which is the interface between the optical field and digitized data - and enables image sensor designs with reduced photosite count. This also enables lower power, lower bandwidth image sensor designs.
• the present invention uses the spatial basis as the measurement basis, exploiting its incoherence with natural images' sparse bases to effectively achieve compressive sensing. This enables the reduction in samples required to reconstruct the image in each color, which, in turn, allows sampling in the three colors to be multiplexed together.
• This system can be augmented with various modifications to the reconstruction technique to improve image quality.
• Correlations between the images in the three basic colors can be leveraged to express the image in a joint basis in which natural images are more sparse than synthetic ones. See Nagesh et al., "Compressive imaging of color images” (IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP), Taipei, Taiwan, April 2009), hereby incorporated by reference in its entirety.
• the joint basis refactors the image in the basic colors in the sparse basis into a part that is common to all colors and color specific parts.
• TV minimization produces better results than just Li minimization.
• TV is a gradient based smoothing functional. See Candes et al, "Practical signal recovery from random projections” (IEEE Trans. Signal Processing, 2005), hereby incorporated by reference in its entirety (IBR).
• Total Variation minimization is implemented in the NESTA solver described in S. Becker, J. Bobin, and E. J. Candes, "NESTA: a fast and accurate first- order method for sparse recovery," In Press, SIAM J. on Imaging Sciences, hereby incorporated by reference in its entirety, currently available at ttp://www.acm.caltech.edu/ ⁇ nesta/ or by request from the authors.
• a conventionally demosaiced image can be used as the starting point for the solvers.
• This demosaicing for randomized CFA patterns can be performed using different methods including ones described in Condat, "Random patterns for color filter arrays with good spectral properties" (Research Report of the IBB, Helmholtz Zentrum Munchen, no. 08-25, Sept. 2008, Kunststoff, Germany), IBR, hereby incorporated by reference in its entirety and in Lukac et al., "Universal demosaicing for imaging pipelines with a RGB color filter array” (Pattern Recognition, vol. 38, pp. 2208-2212, 2005) IBR, hereby incorporated by reference in its entirety.
• Using these as a starting point for image reconstruction can speed up the reconstruction process and, in some implementations, possibly improve the quality of the reconstruction itself.
• the demosaiced image can also be used as a preview of the image prior to reconstruction of the full resolution image.
• CFA patterns can be used. These patterns can be based in different color spaces, i.e., use different sets of basic colors instead of the primary colors red, green and blue.
• CFA patterns with unequal numbers of the basic colors can be used. For example twice as many green sites may be employed as red or blue as is done in the Bayer CFA. Such asymmetries can be useful in reducing noise.
• CFA patterns with certain restrictions on the distribution of the colors such as those discussed in Condat, "Random patterns for color filter arrays with good spectral properties" (Research Report of the IBB, Helmholtz Zentrum Munchen, no. 08-25, Sept. 2008, Kunststoff, Germany), IBR, hereby incorporated by reference in its entirety, may be used. Such restrictions can help avoid excessive clumping of filter elements of identical color while still maintaining the random character of the CFA pattern.
• Another constraint which is statistically linked to natural images is the likelihood of large wavelet coefficients to be at coarser scales. This may be leveraged in the reconstruction process if wavelets are chosen as the sparse basis. Similarly the increased likelihood of large amplitudes to be at lower frequencies in the Fourier transform of natural and most synthetic images can also be leveraged in the reconstruction process.
• Image reconstruction at reduced limiting resolutions of luminance or chrominance or both or of one or more basic colors may result in better image quality.
• the present invention may be used not just for still imaging but for video as well. Besides a trivial extension to multiple frames, algorithms that perform joint reconstruction of multiple frames leveraging their correlation may also be used.
• the present invention may also be used in other situations where multi- spectral image sensor systems are limited by geometric constraints.
• the present invention allows multi- spectral sampling to be folded into smaller sensors requiring smaller apertures without increased acquisition times.
• the present invention provides greater immunity to image quality against dead or malfunctioning sensor elements. This can be used to improve sensor yields: Image sensors with a few dead sensor elements need not be discarded, just characterized.
• the present invention may be used in image scanners.
• the present invention may be used in acquiring multi- spectral images in different number of dimensions including ID and 3D.
• tallf [f ( : , : , 1) f ( : , : ,2) ; f ( : , : ,3) ] ;
• f_noisy f + sigma*abs ( randn ( size ( f ) ) ) ;
• [MAE, MSE, PSNR] Metrics (f , f_noisy) ;
• GPSR_BB (y, A, tau, ...
• fliteredPic uint8 (zeros (size (f ) ) ) ;
• fliteredPic ( : , : , k) uint8 (y ( (k-1 ) *m+l : k*m, : ) ) ;
• tallPic W(theta_debias) ;
• tallPic W(theta) ;
• pic ( : , : , k) tallPic ( (k-1 ) *m+l : k*m, : ) ;
• [MAE, MSE, PSNR] Metrics (f , pic) ;
• pict uint8 (pic) ; imagesc (pict )
• alpha size (A, 1 ) /size (A, 2 ) ;
• alpha2 alpha*alpha
• m size (x, 1 ) /3 ;
• x [idct2 (Lpf ( s ( 1 :m, : ) , 0.65 ) ) ; idct2 ( Lpf ( s (m+1 : 2 *m, : ) , r ) ) ;
• MAE sum (E ) /size (E, 1) ;
• MSE sum(E) /size (E, 1) ;
• PSNR 10*logl0 (255 2/MSE) ;

Landscapes

• Engineering & Computer Science (AREA)
• Multimedia (AREA)
• Signal Processing (AREA)
• Color Television Image Signal Generators (AREA)
• Image Input (AREA)
• Facsimile Heads (AREA)
• Compression Of Band Width Or Redundancy In Fax (AREA)

Priority Applications (4)

Applications Claiming Priority (2)

Publications (2)

Family

ID=44060381

Family Applications (1)

Country Status (7)

Cited By (10)

Families Citing this family (22)

Citations (3)

Family Cites Families (13)

• 2010
  • 2010-11-20 WO PCT/US2010/057541 patent/WO2011063311A2/en active Application Filing
  • 2010-11-20 CA CA2781534A patent/CA2781534A1/en not_active Abandoned
  • 2010-11-20 EP EP10832307A patent/EP2502423A2/en not_active Withdrawn
  • 2010-11-20 US US12/951,043 patent/US8761525B2/en not_active Expired - Fee Related
  • 2010-11-20 KR KR1020127015779A patent/KR20120106762A/ko not_active Application Discontinuation
  • 2010-11-20 CN CN201080060963.0A patent/CN102812709B/zh not_active Expired - Fee Related
  • 2010-11-20 JP JP2012540122A patent/JP2013511924A/ja active Pending

Patent Citations (3)

Also Published As

Similar Documents

Legal Events